Quan Wang

[SenseTime @SenseTime_AI]

𝗢𝗽𝗲𝗻-𝘀𝗼𝘂𝗿𝗰𝗶𝗻𝗴 𝗦𝗲𝗻𝘀𝗲𝗡𝗼𝘃𝗮-𝗩𝗶𝘀𝗶𝗼𝗻-7𝗕-𝗠𝗼𝗧 — 𝗼𝗻𝗲 𝗺𝗼𝗱𝗲𝗹, 𝗮𝗹𝗹 𝗺𝗮𝗷𝗼𝗿 𝘃𝗶𝘀𝗶𝗼𝗻 𝘁𝗮𝘀𝗸𝘀.

Reimagines CV as multimodal generation. Steerable by language or visual prompts. 𝗡𝗼 𝘁𝗮𝘀𝗸-𝘀𝗽𝗲𝗰𝗶𝗳𝗶𝗰 𝗵𝗲𝗮𝗱𝘀 — 𝘀𝘁𝗶𝗹𝗹 𝗦𝗢𝗧𝗔, 𝘀𝘂𝗿𝗽𝗮𝘀𝘀𝗶𝗻𝗴 𝗚𝗼𝗼𝗴𝗹𝗲 𝗗𝗲𝗲𝗽𝗠𝗶𝗻𝗱'𝘀 𝗩𝗶𝘀𝗶𝗼𝗻 𝗕𝗮𝗻𝗮𝗻𝗮 𝗼𝗻 𝘀𝗲𝗴𝗺𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻 𝗮𝗻𝗱 𝗱𝗲𝗻𝘀𝗲 𝗴𝗲𝗼𝗺𝗲𝘁𝗿𝘆.

📊 𝘃𝘀. 𝗩𝗶𝘀𝗶𝗼𝗻 𝗕𝗮𝗻𝗮𝗻𝗮 (𝗚𝗼𝗼𝗴𝗹𝗲 𝗗𝗲𝗲𝗽𝗠𝗶𝗻𝗱):
🔹Referring segmentation (RefCOCOg cIoU): 80.3 vs 73.8
🔹Semantic segmentation (Cityscapes mIoU): 71.2 vs 69.9
🔹Depth estimation (NYUv2 δ1): 98.1 vs 94.8
🔹Surface normal (NYUv2 mean error): 14.4 vs 17.8

Built on a decade of SenseTime's CV leadership — #1 in China's Vision AI market for 10 consecutive years, and named a Tech Innovator in global GenAI CV by Gartner. 𝗩𝗶𝘀𝗶𝗼𝗻, 𝗻𝗼𝘄 𝗻𝗮𝘁𝗶𝘃𝗲 𝘁𝗼 𝗳𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻 𝗺𝗼𝗱𝗲𝗹𝘀. New tasks, defined by instruction — not by new model heads.

[SenseTime @SenseTime_AI]

𝗦𝗲𝗻𝘀𝗲𝗡𝗼𝘃𝗮-𝗩𝗶𝘀𝗶𝗼𝗻-7𝗕-𝗠𝗼𝗧, 𝗳𝘂𝗹𝗹𝘆 𝗼𝗽𝗲𝗻-𝘀𝗼𝘂𝗿𝗰𝗲𝗱: 𝗼𝗻𝗲 𝗺𝗼𝗱𝗲𝗹, 𝗲𝘃𝗲𝗿𝘆 𝗺𝗮𝗷𝗼𝗿 𝘃𝗶𝘀𝗶𝗼𝗻 𝘁𝗮𝘀𝗸 𝗯𝗲𝗹𝗼𝘄:
🔹𝗗𝗲𝘁𝗲𝗰𝘁𝗶𝗼𝗻/𝗢𝗖𝗥/𝗚𝗨𝗜
🔹𝗗𝗲𝗽𝘁𝗵 & 𝗻𝗼𝗿𝗺𝗮𝗹
🔹𝗦𝗲𝗴𝗺𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻
🔹𝗠𝘂𝗹𝘁𝗶-𝘃𝗶𝗲𝘄

It can also define new vision-task variants through natural language — recombining visual capabilities across traditional task boundaries.

Open-sourced: model weights + the SenseNova-Vision Corpus (50M-example subset, plus full toolkit to reproduce the remaining public-source data) for research & development

🤗 HF: https://t.co/dJRnmJ9tPJ
🔧GitHub: https://t.co/CRTDeAigXn
💻 Try the demo: https://t.co/KDSf7xtjnj
📋Technical Report: https://t.co/82wolPMLMR
👾Discord: https://t.co/NdqTQVFANp

[Ziwei Liu @liuziwei7]

🔥Vision as Unified Multimodal Generation🔥

🎯SenseNova-Vision🎯 unifies vision tasks (e.g., detection, keypoints, segmentation, depth, surface normals, point maps, and camera pose) as unified multimodal model (UMM) generation *with SOTA results*

- Code: https://t.co/a1WEmDLVe4

[ModelScope @ModelScope2022]

SenseNova-Vision-7B-MoT is live on ModelScope: a unified multimodal generation model for computer vision. 🚀
🤖 https://t.co/25qQEJ2Mvn
📄 https://t.co/zH10Vcdvcf

📊 Strong against generalist vision models: compared with Youtu-VL, it reports higher COCO detection mAP, Cityscapes mIoU, RefCOCO / RefCOCOg grounding, and NYUv2 depth δ1.

🧩 One model, many output types: structured text for boxes, points, OCR, GUI grounding, keypoints, and camera records; image-like outputs for masks, depth, normals, and point maps.

🧠 No task-specific heads: visual tasks are expressed through text, image, or mixed text-image generation, with outputs decoded back into benchmark formats.

📚 Trained on SenseNova-Vision-Corpus-50M, covering structured visual understanding, dense geometry, segmentation, and multi-view visual geometry.
License: CC BY-NC 4.0.

[Quan Wang @wangquan21]

Thanks for reposting, Songyou! Vision Banana is great and really inspiring! Happy to take this direction one step further by bringing vision closer to foundation models. Looking forward to more discussions!

[Songyou Peng @songyoupeng]

Congrats on the work! Very happy to see this work extended what we showed in Vision Banana to many more vision tasks, and open-source it!

[Songyou Peng @songyoupeng]

Congrats on the work! Very happy to see this work extended what we showed in Vision Banana to many more vision tasks, and open-source it!

[Adina Yakup @AdinaYakup]

SenseNova-Vision 🔥 SenseTime's new model treats all of computer vision as generation

- 7B
- CC BY-NC 4.0 ( non commercial )
- Model/ 50M instruction corpus/ benchmark/ paper/ demo, all open 🫶